Abstract: Certain aspects of the present disclosure are directed to methods and apparatus for configuring a multiply-accumulate (MAC) block in an artificial neural network. A method generally includes receiving, at a neural processing unit comprising one or more logic elements, at least one input associated with a use-case of the neural processing unit; obtaining a set of weights associated with the at least one input; selecting a precision for the set of weights; modifying the set of weights based on the selected precision; and generating an output based, at least in part, on the at least one input, the modified set of weights, and an activation function.

Abstract: A system on chip (SOC) comprising: first memory block and a second memory block; a processing unit coupled to the first memory block and the second memory block; a first power multiplexor disposed between the first memory block and the second memory block and coupled to a first power rail configured to provide an operating voltage to both the first memory block and the second memory block; and enable logic circuitry disposed at a periphery of the SOC away from the first memory block and the second memory block, the enable logic being coupled to control terminals of the first power multiplexor.

Abstract: A system on chip (SOC) includes a power distribution network (PDN) that has two different types of power multiplexers. The first power multiplexer type includes a lower resistance switching logic, and the second type includes a higher resistance switching logic as well as digital logic to provide an enable signal to the first type of power multiplexer. A given first-type power multiplexer may have multiple power multiplexers of the second type in a loop, the loop including communication paths for the enable signal and feeding the enable signal back to an enable input of the first-type power multiplexer.

Abstract: In certain aspects, an apparatus includes a gating circuit having an enable input, a signal input, and an output, wherein the enable input is configured to receive an enable signal. The apparatus also includes a toggle circuit having an output, wherein the toggle circuit is configured to toggle a logic state at the output of the toggle circuit based on the enable signal. The apparatus further includes a multiplexer having a first input, a second input, and an output, wherein the first input of the multiplexer is coupled to the output of the gating circuit, the second input of the multiplexer is coupled to the output of the toggle circuit. The multiplexer is configured to select one of the first input and the second input based on the enable signal, and couple the selected one of the first input and the second input to the output of the multiplexer.

Abstract: A system includes: a first power supply; a second power supply; a headswitch disposed between the first power supply and logic circuitry; an enable driver coupling the second power supply to a control terminal of the headswitch; and a voltage generator operable to adjust a control voltage from the second power supply to the control terminal of the headswitch in response to a first voltage level of the first power supply exceeding a reference voltage level.

Abstract: Certain aspects of the present disclosure are directed to methods and apparatus for configuring a multiply-accumulate (MAC) block in an artificial neural network. A method generally includes receiving, at a neural processing unit comprising one or more logic elements, at least one input associated with a use-case of the neural processing unit; obtaining a set of weights associated with the at least one input; selecting a precision for the set of weights; modifying the set of weights based on the selected precision; and generating an output based, at least in part, on the at least one input, the modified set of weights, and an activation function.

Abstract: In certain aspects, a system comprises a power collapsible logic block, a plurality of retention flip-flops coupled to the power collapsible logic blocks, wherein the plurality of retention flip-flops includes a group of master-slave flip-flops and a group of balloon flip-flops, and a power controller configured to retain states of the group of balloon flip-flops and states of the group of master-slave flip-flops in a first sleep state and to retain the states of the group of balloon flip-flops but not states of the group of master-slave flip-flops in a deep sleep state.

Abstract: Aspects of the disclosure are directed to a digital design with bundled data asynchronous logic and body-biasing tuning. In one aspect, implementation includes establishing a control path between a first controller and a second controller using a handshaking protocol; establishing a data path between a first latch and a second latch using a bundled data technique; executing a first dynamic body biasing tuning by applying a first body bias signal to the control path; executing a second dynamic body biasing tuning by applying a second body bias signal to the data path. The digital design includes a first controller with a control path to connect to a second controller, wherein a first body bias tuning signal tunes body bias in the control path, a first latch with a data path to connect to a second latch, wherein a second body bias tuning signal tunes body bias in the data path.

Abstract: The disclosure generally relates to a memory power reduction scheme that can flexibly transition memory blocks among different power states to reduce power consumption (especially with respect to leakage power) in a manner that balances tradeoffs between reduced power consumption and performance impacts. For example, according to various aspects, individual memory blocks may be associated with an access-dependent age, whereby memory blocks that are not accessed may be periodically aged. As such, in response to the age associated with a memory block crossing an appropriate threshold, the memory block may be transitioned to a power state that generally consumes less leakage power and has a larger performance penalty. Furthermore, one or more performance-related criteria may be defined with certain memory blocks to prevent and/or automatically trigger a transition to another power state.

Abstract: In certain aspects, an apparatus includes a first plurality of power switch devices. Each of the first plurality of power switch devices includes a delay line having a programmable time delay, and a power switch coupled between a supply rail and a circuit block, wherein the power switch has a control input coupled to the delay line. The apparatus also includes a switch manager configured to program the time delays of the delay lines in the first plurality of power switch devices based on a number of active circuit blocks in a system.

Abstract: Electronic devices employing adiabatic logic circuits with wireless charging are disclosed. In one aspect, an electronic device is provided. The electronic device includes a power circuit employing an alternating current (AC) coupler circuit configured to receive a wireless AC signal and generate a wired AC signal based on the wireless AC signal. The power circuit includes a power output configured to provide an AC power signal based on the wired AC signal generated by the AC coupler circuit. The AC power signal is generated based on the wireless charging capability of the AC coupler circuit. The electronic device employs a digital logic system that includes a power rail electrically coupled to an adiabatic logic circuit. The AC power signal is provided to the power rail to provide power to the adiabatic logic circuit. Wirelessly charging the adiabatic logic circuit consumes less power than conventional non-wireless charging circuitry.

Abstract: Adaptive power regulation methods and systems are disclosed. In one aspect, one or more process sensors for memory elements are provided, which report information relating to inherent speed characteristics of sub-elements within the memory elements. Based on this reported information, a controller ascertains an appropriate power level to insure a proper data retention voltage (DRV) is applied on voltage rails by a power management unit (PMU) circuit. By using the proper DRV based on the speed characteristics of the sub-elements within the memory elements, power conservation is achieved.

Abstract: Aspects of the disclosure are directed to a digital design with bundled data asynchronous logic and body-biasing tuning. In one aspect, implementation includes establishing a control path between a first controller and a second controller using a handshaking protocol; establishing a data path between a first latch and a second latch using a bundled data technique; executing a first dynamic body biasing tuning by applying a first body bias signal to the control path; executing a second dynamic body biasing tuning by applying a second body bias signal to the data path. The digital design includes a first controller with a control path to connect to a second controller, wherein a first body bias tuning signal tunes body bias in the control path, a first latch with a data path to connect to a second latch, wherein a second body bias tuning signal tunes body bias in the data path.

Abstract: Methods and apparatus for identifying a process corner are provided. Provided is an exemplary method for identifying a process corner of an integrated circuit (IC). The IC has a first asymmetrical ring oscillator (ARO1) including pull-up transistors that have a low threshold voltage (LVT) and pull-down transistors that have a regular threshold voltage (RVT), and has a second asymmetrical ring oscillator (ARO2) including pull-up transistors that have an RVT and pull-down transistors having an LVT. The exemplary method includes applying an ultra-low power supply voltage to the ARO1 and the ARO2 that causes the integrated circuit to operate near a verge of malfunction, measuring an output frequency of the ARO1, measuring an output frequency of the ARO2, calculating a calculated ratio of the output frequency of the ARO1 and the output frequency of the ARO2, and comparing the calculated ratio to a fiduciary ratio to identify the process corner.

Abstract: Clock distribution schemes with wide operating voltage ranges are disclosed. In one aspect, an operating voltage level or condition within a computing device is sensed. In a first voltage condition, delay elements are used within a clock tree to minimize clock skew. In a second voltage condition, one or more delay and/or clocked elements are bypassed to minimize clock skew at the second voltage condition. In addition to controlling clock skew, power may be conserved by depowering the bypassed elements. Controlling clock skew in this fashion improves operation of a computing device that includes the clock tree and may improve battery life.

Abstract: A device includes a first set of storage elements, a second set of storage elements, and a bias circuit configured to generate a test bias signal to bias the first set of storage elements and the second set of storage elements. The device further includes a sensor circuit configured to receive a first signal from at least a first storage element of the first set of storage elements in response to the test bias signal and to receive a second signal from at least a second storage element of the second set of storage elements in response to the test bias signal. The sensor circuit is further configured to generate a third signal having a delay characteristic indicating a wear difference between the first storage element and the second storage element.

Abstract: The disclosure generally relates to a memory power reduction scheme that can flexibly transition memory blocks among different power states to reduce power consumption (especially with respect to leakage power) in a manner that balances tradeoffs between reduced power consumption and performance impacts. For example, according to various aspects, individual memory blocks may be associated with an access-dependent age, whereby memory blocks that are not accessed may be periodically aged. As such, in response to the age associated with a memory block crossing an appropriate threshold, the memory block may be transitioned to a power state that generally consumes less leakage power and has a larger performance penalty. Furthermore, one or more performance-related criteria may be defined with certain memory blocks to prevent and/or automatically trigger a transition to another power state.

Abstract: Adaptive power regulation methods and systems are disclosed. In one aspect, one or more process sensors for memory elements are provided, which report information relating to inherent speed characteristics of sub-elements within the memory elements. Based on this reported information, a controller ascertains an appropriate power level to insure a proper data retention voltage (DRV) is applied on voltage rails by a power management unit (PMU) circuit. By using the proper DRV based on the speed characteristics of the sub-elements within the memory elements, power conservation is achieved.

Abstract: Electronic devices employing adiabatic logic circuits with wireless charging are disclosed. In one aspect, an electronic device is provided. The electronic device includes a power circuit employing an alternating current (AC) coupler circuit configured to receive a wireless AC signal and generate a wired AC signal based on the wireless AC signal. The power circuit includes a power output configured to provide an AC power signal based on the wired AC signal generated by the AC coupler circuit. The AC power signal is generated based on the wireless charging capability of the AC coupler circuit. The electronic device employs a digital logic system that includes a power rail electrically coupled to an adiabatic logic circuit. The AC power signal is provided to the power rail to provide power to the adiabatic logic circuit. Wirelessly charging the adiabatic logic circuit consumes less power than conventional non-wireless charging circuitry.

Abstract: A switchable supply network for powering multiple digital islands. In one embodiment, a first digital island includes a first power collapsible circuit and a first retention circuit, and a second digital island includes a second power collapsible circuit and a second retention circuit. In a normal mode of operation, the first digital island is provided a first supply voltage and a second digital island is provided a second supply voltage higher than the first supply voltage. In a transition mode the second power collapsible circuit is powered down and the second supply voltage is lowered and provided to the second retention circuit. When the second supply voltage falls below the first supply voltage, the first power collapsible circuit is powered down. The second supply voltage is now provided only to the retention circuits, and is furthered lowered in a retention mode to a final retention voltage.